OpenAI announced the addition of two advanced security measures for ChatGPT to address the growing risk of prompt injection attacks when AI systems connect to websites and external applications. These attacks can be exploited by third parties to trick the AI into executing malicious commands or leaking sensitive information. Both new measures are built upon the platform's existing sandbox mechanism, URL data leakage protection, and other multi-layered security systems.

image.png

The first protective measure introduced is Lockdown Mode (Lockdown Mode), an optional advanced setting for users with high security needs, not intended for general users. It reduces the risk of data leakage by strictly limiting how ChatGPT interacts with external systems. This mode deterministically disables or limits tools that are vulnerable to exploitation, such as web browsing, which will only access cached content, and features without strong security guarantees will be directly disabled.

This mode currently applies to ChatGPT Enterprise, Edu, healthcare, and teacher versions. Administrators can create roles in the dashboard settings to enable it and can finely control which external applications and specific operations are allowed under Lockdown Mode; OpenAI also plans to open this mode to consumer users in the coming months. Additionally, the Compliance API Logs Platform can help administrators carry out regulatory audits on application usage and data sharing.

The second measure is to uniformly label functions with higher security risks in ChatGPT, ChatGPT Atlas, and Codex with the **"Elevated Risk (Increased Risk)"** tag, standardizing the way risks are alerted.

OpenAI stated that some network-related capabilities, although they can enhance the practicality of AI, still carry risks that have not been fully addressed by industry protection measures. The label allows users to clearly understand the risks and decide for themselves whether to use them, especially in scenarios involving private data. For example, when developers enable network access for Codex, the label will appear on the interface, clearly informing them of changes after activation, potential risks, and suitable scenarios, along with risk mitigation suggestions.